Manufactured Truth: Has the Age of AI Deception Begun?

In August 2026, an American think tank called the “Hanover Institute for Public Policy” appeared online. It had no address, no names of researchers, and did not appear to exist as a legal entity. Yet, over the course of just nine days, it published 124 reports totaling more than 560,000 words. Its topics were highly sensitive: Is Israel committing genocide in Gaza? Did it expel Palestinians from their land? Did it deliberately starve Palestinians? And is anti-Zionism inherently antisemitic? Its publications presented the Israeli position on these issues in the form of seemingly well-documented research, before The Guardian revealed that the website was part of an effort funded by the Israeli government through intermediaries, and that it had used a commercial platform designed to increase the likelihood that its content would be cited by AI systems such as ChatGPT, Claude, Gemini, and Perplexity. The U.S. company behind the project said its work was intended to place accurate, well-documented facts into the public record and counter misinformation about Israel. Perhaps the significance of the story lies less in the political position it seeks to influence than in the identity of the reader it is trying to reach. Some people have begun writing online not because they want to persuade you first, but because they want to persuade the machine that will eventually answer your questions.

For more than twenty years, we have known of attempts to manipulate search engines. Alongside legitimate search-engine optimization practices, there have been efforts to understand ranking signals and exploit them. Yet humans ultimately remained in front of a list of results: they chose a page, saw the website's name, and read the text. Answer engines, by contrast, can search, select, read, summarize, and then return what they have found in a single answer. That is why the shift from Search Engine Optimization (SEO) to Generative Engine Optimization (GEO) may represent more than simply the marketing industry's transition from one technique to another. The competition is no longer only over who ranks first, but over what gets into the answer in the first place.

This is where something fundamental changes in the nature of disinformation itself. An old-fashioned lie needed a human being to believe it, and propaganda needed an audience to see it. Today, however, content can be produced without its creator particularly caring whether we read it. It may be enough for that content to be discoverable in ways that increase the likelihood of AI systems finding it. In May 2026, researchers T. Tony Ke, Chenxi Liao, and Xiaoyan Xu published a paper on the concept of a “Synthetic Corpus,” examining the possibility of populating the digital environment with synthetic materials, such as reviews and articles, in order to influence what AI systems subsequently present to users. The study does not establish that every model can be manipulated in this way, but it reveals that the contest is no longer only about the message itself. It is also about the environment in which the machine searches for that message. We are no longer merely fabricating the lie; we are potentially fabricating the environment that makes the lie appear credible.

A single claim can be challenged, and a single website can be exposed as biased. But what happens when a system searches and finds dozens of pages repeating the same claim, under different names and in different styles, appearing to come from independent organizations? What happens when the appearance of consensus can be manufactured before the machine is even asked what the truth is? This does not mean that models will treat every webpage as an independent source, or that repetition alone is enough to deceive them. A recent review covering 45 studies in the field of GEO found that the evidence remains inconsistent, and that there is still no proven formula capable of producing a stable, long-term effect across platforms. But we are no longer dealing with a purely theoretical possibility. Some actors are already attempting to influence the material machines will encounter before that material reaches us.

Perhaps the most vulnerable space is not the subject about which much has already been written, but the one about which insufficient material exists. A major historical event surrounded by thousands of books, documents, and archival records is difficult to reshape simply by adding dozens of webpages. But what about a local incident documented by only a few sources, the history of a community whose records have not been digitized, or an issue that has received little attention from major institutions? In such cases, those seeking to influence the narrative may not need to defeat the truth; they may simply need to find the gaps that the truth has not yet reached. Those seeking to distort history may not need to erase it; sometimes, it may be enough to write more about it than the people who actually lived it. This is a question that deserves attention in the Arab world as well—not because there is sufficient evidence to claim that Arabic-language content has already become a victim of this phenomenon, but because the digital presence of Arab knowledge is no longer relevant only to human readers. A book that has never been digitized, a document that remains in an archive, a study that is difficult to access, or a local narrative that has never been properly documented may also be absent from some of the pathways machines use to construct their answers. In the age of search engines, what was absent from the internet was simply less visible. In the age of answer engines, it may become less present in the narrative the machine constructs about the world.

The issue becomes even more sensitive when we consider what humans do after an answer appears. In an analysis published by the Pew Research Center in 2025 of the behavior of 900 U.S. users, covering 68,879 Google searches, clicks on traditional search results fell from 15% when no AI-generated summary appeared to 8% when one did. Meanwhile, users clicked on links cited by the AI-generated summary in only about 1% of visits. These figures do not represent all AI users or every answer engine, but they say something difficult to ignore: the machine reads the sources more than the human does, while the human reads what the machine has written about them.

Yet it would be too easy to turn the human into a victim and the machine into the accused. The same artificial intelligence that has introduced a new problem into our relationship with truth has also given us a new tool for examining it. A user can take an answer generated by one model and submit it to another—not to ask it the same question, but to ask it to treat the answer as if it were material submitted for peer review: to revisit the literature, examine the quality of the studies, search for opposing views, compare the strength of the evidence, and ask whether the conclusion is actually consistent with what the research says. This practice is increasingly reflected in research exploring the use of AI models for criticism, evaluation, and verification, although the available evidence does not yet justify treating them as a reliable substitute for human review. Here lies the paradox: AI has lowered the cost of producing disinformation, but it has also lowered the cost of doubting it. The existence of sources, in either case, is not enough. References may all be genuine while the answer itself is misleading because it selectively cites supportive studies while ignoring contradictory ones, relies on a weak study while overlooking a stronger systematic review, or reports an accurate finding and then builds an unsupported conclusion upon it. We may therefore need to move from asking, “Does this claim have a source?” to a more difficult question: What does the body of evidence say, how strong is it, and which interpretations are supported by the strongest research?

Yet this solution brings us, strangely enough, back to the same problem. If we take an answer generated by one model and ask a second model to challenge it and return to the literature, its success ultimately depends on the knowledge environment in which it searches. If that environment is filled with weak, biased, or synthetic material, the model may find ten sources without realizing that all ten ultimately trace back to a single claim, or that their quantity does not equal the strength of one more rigorous study. The question we need to teach machines is the same one we should have learned ourselves: Does the number of sources indicate stronger evidence, or merely more repetition? The research environment itself is not entirely immune either. In a 2026 study, researchers analyzed roughly 111 million references appearing in 2.5 million papers on arXiv, bioRxiv, SSRN, and PubMed Central, and conservatively estimated that 146,932 references cited in papers from 2025 alone did not actually exist, with a clear increase following the spread of large language models. Here a difficult feedback loop emerges: a machine generates information or a reference, it finds its way into the knowledge environment, and another machine then searches that same environment to verify a new piece of information. What we therefore need is not only an independent critic, but also an independent and trustworthy evidence base.

The paradox is that the internet, which lowered the cost of publishing, is now being joined by AI, which is lowering the cost of filling it. Producing hundreds of thousands of words once required writers, time, and money. Today, vast quantities of text can be generated faster and more cheaply. The old question—“Who can publish?”—may therefore be giving way to another: “Who can fill the gaps?” If producing material that looks like knowledge becomes cheaper than examining and verifying its origins, we are not facing merely an abundance of information, but an abundance of the appearance of knowledge. A single lie is a content problem; manufacturing an environment in which that lie appears to be true is a problem with the architecture of knowledge itself.

This is why the danger of the think tank story with which we began does not lie in proving that someone has successfully rewritten what machines know. There is, at present, no evidence sufficient to draw that conclusion. Its significance lies instead in the fact that the idea is now operationally possible, and that some actors have already begun producing material whose success is measured not only by the number of humans who read it, but by the likelihood that it will enter the pathways AI systems use when searching for answers. This makes the expression “writing for machines” more than a marketing term, because some content may no longer have a human being as its primary target. The internet, which was once a space where people competed for our attention, may become a space where they also compete for the attention of the machines we ask to know the world on our behalf.

For years, we have tried to teach people not to believe everything they see on the internet. We may now be entering an era in which we need to protect the environment in which machines search when they attempt to know on our behalf, to document what we know before those who can produce text fastest fill the gaps, and to remember that dozens of links do not create truth, just as repetition does not turn a claim into a consensus. Truth does not become stronger because it has been stated more times; it becomes stronger because the evidence leading to it survives every attempt to test it. Perhaps that is why the most important question is no longer, “Can AI lie?” A machine can make mistakes, hallucinate, and correct itself. But what is unfolding now confronts us with a more complicated question, because the problem is no longer only what the machine tells us, but what we put into the world for it to find when it searches. If AI reads the internet to tell us what is true, who will write the truth that AI finds there?